The Trump administration is actively considering mechanisms to restrict US access to Chinese frontier AI models, according to Axios. The Commerce Department last year considered adding multiple Chinese AI labs to its Entity List, which would effectively cut off US access without a license.
Administration officials who favored keeping regulation from stifling innovation killed those earlier efforts, Axios reports. But Moonshot AI’s launch of Kimi K3 last week has reignited the push. Chinese frontier models are now demonstrably competitive with US systems on multiple benchmarks, narrowing a performance gap that previously gave American labs a structural advantage.
Neither the White House nor the Commerce Department responded to Axios’s requests for comment.
The Policy Mechanism
The Entity List is the Commerce Department’s primary tool for restricting US business relationships with foreign entities deemed national security risks. Adding Chinese AI labs would not require an outright ban on their models. It would instead create a licensing barrier that, in practice, would push US enterprises away from Chinese platforms entirely.
The administration would not need to impose a formal prohibition to achieve this outcome. The compliance burden of Entity List designation typically causes US companies to drop affected vendors preemptively rather than navigate the licensing process.
The Competition Argument
David Sacks, an outside White House AI adviser, pushed back publicly on Sunday. “We are at a critical inflection point in AI policy,” Sacks wrote on X. “The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition.”
The concern from pro-competition voices is straightforward: restricting Chinese models removes pricing and performance pressure on OpenAI and Anthropic, locking US enterprise customers into a two-vendor market. Chinese open-weight models like Kimi K3 and Qwen 3.8 have been gaining adoption precisely because they offer competitive performance at lower cost.
According to a UK AI Security Institute benchmark published this month, open-weight models now trail frontier closed models by only four to seven months on autonomous cyberattack tasks, down from six to ten months in 2025. The performance convergence is real, and it extends beyond security benchmarks into general reasoning and code generation.
Anthropic and OpenAI’s Position
Both companies have long advocated for safety regimes that in some cases include model licensing requirements. Those advocacy efforts were previously abstract, focused on hypothetical risks from frontier capabilities. The Entity List consideration makes them concrete: safety-motivated licensing would now serve a competitive function, shielding US closed-source labs from foreign open-weight alternatives.
Axios notes that powerful US alternatives remain limited compared to cheaper Chinese models in several practical deployment scenarios.
The Geopolitical Shift
The policy discussion marks a strategic pivot. Rather than competing on model performance, the US would be competing on access control: who gets to use which models in which markets. AI infrastructure joins semiconductors and compute on the list of technologies where geopolitical competition is now primarily a regulatory contest rather than a technical one.
Chinese leader Xi Jinping is reportedly feeling emboldened in the AI race, and pressure is mounting on the US industry to respond. For enterprise teams that adopted Chinese open-weight models for cost or flexibility reasons, the Entity List threat introduces supply chain risk that may force re-evaluation regardless of whether a formal restriction materializes.